Data Package Metadata   View Summary

The Lake Sunapee LSPA instrumented buoy was deployed on 27 August 2007 and anchored at ~15 m depth near Loon Island at 43.39° N, 72.06° W for the study period (August 2007 – December 2008). The buoy is equipped with a LI-COR LI-190SB photosynthetically active radiation (PAR) sensor (2007-present) and a wind speed and wind direction anemometer (WMT50 Vaisala). The wind speed and direction anemometer is located approximately 1.7 meters above the water surface. In addition, the buoy is equipped with a thermistor chain of TempLine thermistors (Apprise Technology). A Zebra-Tech D-Opto optical dissolved oxygen (DO) and water temperature sensor was deployed at 1 meter for the study period and remained below the ice during the winter period. Data were recorded from all sensors every 10 minutes. These data have been QAQC’d to remove obviously errant readings, highly suspicious readings, and artifacts of buoy maintenance. Data were collated and cleaned using R (v3.5.3). Artifacts of buoy maintenance and errant readings were recoded to ‘NA’ from the raw data.

Dissolved oxygen (DO) sensor readings drifted from the installation date. We corrected the DO data for drift as follows: We assumed the first 60 days from installation were calibrated correctly (as per most manufacturer’s suggestions). Following those 60 days, we assumed linear drift for the sensor. From down sampled DO data, we fit a time series model to the daily time series of DO saturation (%) using the auto.arima function from the R forecast package (Hyndman and Khandakar 2008) with a linear regression term to account for daily drift starting on the 61st day and a constant offset. A first-order autoregressive model was selected with four moving average terms (ARIMA, [1, 0, 4]) and normally distributed residuals (N [0, σ2 = 2.34]). The DO saturation was corrected for drift using the modeled drift rate of 0.034% d-1, lowered by the modeled offset of 17%, and converted back to DO concentration. The corrected DO concentration was used for all subsequent metabolism modeling.

The method we used to calculate lake metabolism estimates is similar to the maximum likelihood estimation method in LakeMetabolizer, which is also used in Solomon et al. (2013) and Hanson et al. (2008). We used the drift-corrected dissolved oxygen concentration data, water temperature profiles to calculate the mixed layer depth, above-water photosynthetically active radiation (PAR) to calculate irradiance, and wind speed data to calculate the gas flux coefficient from the gas exchange piston velocity (k) according to the empirical model from Cole and Caraco (1998) and similar to the k.cole.base function in LakeMetabolizer. We used an inverse modeling technique to fit daily GPP and R parameters with a maximum likelihood method and the Nelder-Mead negative log-likelihood optimization algorithm.

We modified the typical open-water techniques to estimate metabolism used in Solomon et al. (2013) because the diel changes in DO concentrations can be very small in oligotrophic lakes, following the methods described in Richardson et al. (2016). The first modification was that the initial modeled DO concentration was calculated from a mean of the first hour of data one hour before sunrise each day instead of using a single-point estimate. This modification was made to reduce the effect of a potentially erroneous single-point DO measurement due to non-metabolism related processes that create noise in the time-series. Second, in order to include dawn periods prior to sunrise in the daily GPP and R estimates, we calculated a solar day that included one hour before sunrise and one hour after sunset for a total of a 26-hour period. The third modification was similar to Richardson et al. (2016), in which we noticed that there was high variability in DO concentration after sunrise for 1-5 hours throughout a majority of the dataset. We are not sure what caused this signal but after comparing a removal of 1, 2, 3, 4, 5, and 6 hours after sunrise, we chose to remove a 5-hr period for all days because that best fit the observed diel curve (following Richardson et al. 2016). Metabolic rates were calculated with the remaining 21 hours of data. We followed Richardson et al. (2016) to model metabolism on days with open water, but we assumed no atmospheric exchange on ice-covered days (6 Dec 2007–25 April 2008).

Following the modifications to the metabolism model, good model fits were determined by examining the observed and modeled DO data for each day in our time series. We found that removing the first 5 hours of data beginning at 1 hour before sunrise resulted in more biologically meaningful estimates of GPP and R with values >0.001 mg O2 L-1 day-1. However, we did still have some days with daily estimates <0.001 mg O2 L-1 day-1, which were removed. We then followed a similar protocol to Richardson et al. (2016) to conservatively exclude days with unrealistic model estimates for GPP and R. We used plots of observed and modeled diel DO curves, DO residuals, wind speed, and PAR data to visually assess the model fit for each day as acceptable or unacceptable. Three of the authors (JAB, DCR, NKW) independently examined each model fit and the fit statistics (AICc, R2, and residual sum of squares). We compiled the assessments, and if two or more authors determined the model fit was unacceptable, the metabolism estimates from that day were excluded from all subsequent analyses. We used this visual assessment technique because the fit statistics were not a reliable indicator of days with adequate model fits, likely due to the large number of data points (given that the DO sensor collected data every 10 minutes) generating each fit. This method resulted in conservative estimates of GPP and R for this dataset because days with poor model fits would have likely been included if numeric criteria alone were used.

We converted epilimnetic volumetric based rates of GPP and R to areal rates by multiplying by the seasonal thermocline depth when the lake was stratified or the maximum lake depth at the buoy site when the lake was not stratified. The seasonal thermocline depth was calculated using methods similar to Read et al. (2011) using a minimum density threshold of 0.1 kg m3 to find the depth with the maximum density difference and determined using R code available upon request. The maximum lake depth was 15 m for all time periods when the lake was not stratified, except during the winter when the ice depth was 1 m, so the maximum lake depth was 14 m. Daily NEP was calculated as GPP-R for each day with reasonable estimates and adequate model fits as described above.

For the environmental correlates analysis, we used daily sums of precipitation and snow depth data collected at the Newport, NH, USA (Lat: 43.3772° N, Long: 72.1812° W) NOAA NCDC weather station (Station #: GHCND:USC00275868, URL: http://www.ncdc.noaa.gov/cdo-web/search). From the sensors on the Lake Sunapee buoy, we down sampled the 10-minute data to calculate the solar day mean and coefficient of variation (CV) of water temperature collected at the DO sensor, solar day sum of above lake PAR, and the solar day maximum and CV of wind speed (DOI for Sunapee meteorological data: 10.6073/pasta/175cfde22aaeee7349285d2c9fd298d9). From the high-frequency water temperature profile data, solar day max and CV summaries of 10-minute Schmidt stability estimates were derived using the rLakeAnalyzer function (Winslow et al. 2018) following methods in Read et al. (2011).